Prompting the Muse: How Writers Are Using AI as a Co-Author in Creative Writing Programs

A novelist opens ChatGPT next to her manuscript in a peaceful Seattle apartment. She types a prompt to get unstuck, not to create her next chapter. She desires diversity. alternatives. Not precisely the solution, but enough friction to identify what she doesn’t want, which will ultimately reveal her actions. After reading the output, she returns to her own document without using any of it. The scene is written thirty minutes later. Was it dishonest? She doesn’t believe that. The researchers who watched writers for nine months don’t seem to enjoy her work either.

One of the more cautious studies to come out of the ongoing and somewhat nervous discussion about AI and creative writing is the University of Washington study, which was published in late 2024 and was led by Alicia Guo and colleagues like Jeffrey Heer and Amy Zhang. Eighteen authors of poetry, fiction, and creative nonfiction were observed and interviewed. The discovery that garnered the most attention—and maybe merits more attention than it did—is that the writers who used AI the most successfully did not offload the most work. They were the ones who used it most purposefully and the least.

That distinction is important. Two groups have dominated the public discourse on AI and creative writing: proponents who view it as a revolutionary tool for productivity and detractors who view its use as a betrayal of the craft. Neither of these descriptions applied to the authors in this study. They had carefully considered which aspects of the writing process felt like theirs and which felt like friction to be managed, and they were protective of their work—sometimes almost defensively so. For the majority of them, writing prose was their responsibility. AI gained traction in research, brainstorming dead ends, and unlocking obstacles at three in the morning.

Although it posed the question in a different way, Glenn Kleiman’s earlier Stanford working paper from 2022 came to a similar set of conclusions. Although the boundaries have become much more hazy since GPT-3 was the most potent system available, his four roles for AI in writing—editor, co-author, ghostwriter, and muse—still hold up as a taxonomy. For good reason, publishers and literary magazines are most anxious about the ghostwriter position. There was an immediate and severe backlash when NaNoWriMo announced that it would permit the use of AI in its month-long novel-writing challenge. After being overrun with AI-generated content, a number of literary magazines temporarily stopped accepting submissions. Limiting the use of AI in screenwriting was one of the specific demands made during the 2023 Writers Guild of America strike. These issues have not yet been settled by the creative community, and it is still unclear if they will be or if circumstances will force a decision instead of consensus.

The University of Washington study demonstrates that authors who have successfully incorporated AI into their work on a regular basis, rather than as a novelty, have created something more complex than approval or disapproval. They have created their own policies. They set rules for themselves regarding what AI can and cannot touch. In the study, one writer completely relied on AI as a muse, giving it questions and sparks but never allowing it to produce prose. Another used it as a ghostwriter to write in his own voice, but he strongly objected to any creative input. The taxonomy maps to actual choices made at actual times during actual writing sessions; it is not abstract.

Ted Chiang made an important point in his essay “Why AI Isn’t Going to Make Art,” which became a focal point for the entire discussion. Ten thousand options are needed for a short story. One hundred is needed for a prompt. One could argue that what is lost in the space between those figures is the purpose of writing, not only the final product but also the thought that goes into it. This was intuitively understood by the authors of the Seattle study. The ten thousand options were not intended to be automated. Something had momentarily blocked their path as they attempted to resume making them.

Observing this field from the outside, it seems as though the loudest debates on both sides are overshadowing the real experience of writers using these tools in silence, developing techniques that don’t neatly fit into anyone’s preferred narrative. The creative writing programs that will benefit students the most over the next ten years are likely to be the ones that are open to having that honest discussion. Instead of viewing AI as either the enemy or the future of writing, they should view it as a tool whose usefulness solely depends on the writer’s goals and level of self-awareness.